Seth Karten

Seth Karten

@sethkarten.ai

Autonomous Agents | Research @ Prime Intellect | PhD @ Princeton | Prev: CMU, Waymo | NSF GRFP Fellow https://sethkarten.ai/

636 followers1.4K following243 posts

Seth Karten reposted

Seth Karten
@sethkarten.ai

Over two weeks, Prime Agent orchestrated a swarm of over 2,000 agents to rewrite itself in Rust. With 10,000+ sandboxes, 200B+ GLM-5.3 tokens, and 16,000 agent-to-agent messages.. Prime Agent now reaches usable input ~13× faster and uses 83% less startup memory than before.

Seth Karten
@sethkarten.ai

Over two weeks, Prime Agent orchestrated a swarm of over 2,000 agents to rewrite itself in Rust. With 10,000+ sandboxes, 200B+ GLM-5.3 tokens, and 16,000 agent-to-agent messages.. Prime Agent now reaches usable input ~13× faster and uses 83% less startup memory than before.

View full thread
Seth Karten
@sethkarten.ai

Over two weeks, Prime Agent orchestrated a swarm of over 2,000 agents to rewrite itself in Rust. With 10,000+ sandboxes, 200B+ GLM-5.3 tokens, and 16,000 agent-to-agent messages.. Prime Agent now reaches usable input ~13× faster and uses 83% less startup memory than before.

View full thread
Seth Karten
@sethkarten.ai

Across the six harnesses we benchmarked, Prime Agent Rust had the lowest time to usable input, startup memory, and installed size.

Seth Karten
@sethkarten.ai

We’re turning the workflow behind this rewrite into reusable state machines in Prime Agent, so you can run your own teams of agents through implementation, review, and verification.

Seth Karten
@sethkarten.ai

Over two weeks, Prime Agent orchestrated a swarm of over 2,000 agents to rewrite itself in Rust. With 10,000+ sandboxes, 200B+ GLM-5.3 tokens, and 16,000 agent-to-agent messages.. Prime Agent now reaches usable input ~13× faster and uses 83% less startup memory than before.

View full thread
Seth Karten
@sethkarten.ai

A root agent divided the rewrite into tasks and tracked their dependencies. Each task ran through its own state machine: Planner → Implementer → Reviewer → Verifier

Seth Karten
@sethkarten.ai

With feature parity achieved, we gave Prime Agent a runtime benchmark suite and let it hillclimb its own performance.

Seth Karten
@sethkarten.ai

Over two weeks, Prime Agent orchestrated a swarm of over 2,000 agents to rewrite itself in Rust. With 10,000+ sandboxes, 200B+ GLM-5.3 tokens, and 16,000 agent-to-agent messages.. Prime Agent now reaches usable input ~13× faster and uses 83% less startup memory than before.

Seth Karten
@sethkarten.ai

To preserve the TypeScript version’s behavior, we gave the agents objective parity checks: • Diff terminal frames • Compare session transcripts and model requests • Check daemon protocol messages • Audit every feature They could see where behavior diverged and fix it before changes merged.

Seth Karten
@sethkarten.ai

Attending COLM 2026? Come say hi to talk about self-improving agents and economic multi-agent alignment Looking forward to chatting with old and new friends!

Seth Karten
@sethkarten.ai

We wrote Agent Bazaar back in May around a future where agents act on behalf of users and increasingly participate directly in marketplaces like Amazon and eBay. We introduced Economic Alignment to study what happens as agents become economic actors.

Seth Karten
@sethkarten.ai

The risks go beyond whether any single agent is aligned with its user. Agents can collectively crash markets even when each individual agent is behaving rationally, and users can get scammed by autonomous sellers. Meta Muse is starting to make this future real pretty quickly.

Seth Karten
@sethkarten.ai

This is exactly the kind of setting we built Agent Bazaar to study, and I think these questions are only going to become more important as agents start transacting and negotiating on behalf of users. I'll be presenting Agent Bazaar at COLM Wed Oct 7. arxiv.org/abs/2605.17698

arxiv.orgAgent Bazaar: Enabling Economic Alignment in Multi-Agent MarketplacesThe deployment of Large Language Models (LLMs) as autonomous economic agents introduces systemic risks that extend beyond individual capability failures. As agents transition to directly interacting w...
Seth Karten
@sethkarten.ai

prime agent v0.9.6 is out: ◆ Support for GPT-6 Sol, Opus 5.5, and Grok 4.7 ◆ /mcp plugin catalog with one-click connections to Linear, Notion, Posthog, Stripe, and 60+ more services ◆ Huge perf and reliability pass 🫡 Lots more coming soon :)

Seth Karten
@sethkarten.ai

prime agent v0.9.5 we fixed a lot of bugs and, of course, we had prime agent feature its favorite updates it picked our perf work. then it created the video itself.

Seth Karten
@sethkarten.ai

Quick update: we’re adding a few incentives for contributors. Everyone who submits data will be acknowledged in the dataset release, and the full dataset will be open sourced. If you get especially involved in collection and data processing/curation, there may also be author opportunities

Seth Karten@sethkarten.ai

My group at Princeton is collecting crowdsourced, action-labeled game data for research on AI agents, starting with Pokémon Emerald. The goal is to build a large open-source dataset of action-labeled gaming trajectories. Sign up to contribute here: docs.google.com/forms/d/e/1F...

docs.google.comPlay Pokémon Emerald, Help Train a World Model.We're building generalizable world model agents, AIs that learn to understand and simulate games from video. We need your help to gather gameplay recordings to generate a dataset which can then teach ...

Seth Karten reposted

prolepses.bsky.social
@prolepses.bsky.social

Reposting with a direct link right to where @sethkarten.ai breaks down Prime Agent. The TLDR is that Prime Agent is "Jupyter notebooks for your agents", the video is absolutely worth the watch: youtu.be/n9xKblqyQ28?...

youtu.beWhy The Harness Matters More Than The Model | YC Paper ClubYouTube video by Y Combinator
Chris Patil@donotgogently.bsky.social

If you’re interested in Prime Agent by @primeintellect.bsky.social , Seth Karten (researcher at Prime, author of the Prime Agent paper) gave a nice talk on the harness at YCombinator the other day youtu.be/n9xKblqyQ28?...

youtu.beWhy The Harness Matters More Than The Model | YC Paper ClubYouTube video by Y Combinator
Seth Karten
@sethkarten.ai

My group at Princeton is collecting crowdsourced, action-labeled game data for research on AI agents, starting with Pokémon Emerald. The goal is to build a large open-source dataset of action-labeled gaming trajectories. Sign up to contribute here: docs.google.com/forms/d/e/1F...

docs.google.comPlay Pokémon Emerald, Help Train a World Model.We're building generalizable world model agents, AIs that learn to understand and simulate games from video. We need your help to gather gameplay recordings to generate a dataset which can then teach ...
Seth Karten
@sethkarten.ai

Wow, three papers in CoLM 2026... Here I come San Francisco! These papers predicted some early trends in multi-agent safety & economic envs, automatic RL env creation, and PPO for VLMs/LLMs 🧵

View full thread
Seth Karten
@sethkarten.ai

Wow, three papers in CoLM 2026... Here I come San Francisco! These papers predicted some early trends in multi-agent safety & economic envs, automatic RL env creation, and PPO for VLMs/LLMs 🧵

Seth Karten
@sethkarten.ai

New blog applying Continual Harness to ARC-AGI-3. The heavy test-time learning required by the benchmark pushes agents to form an internal world model of the rules and mechanics that updates with new evidence. Continual Harness scored 20.54%. sethkarten.substack.com/p/continual-...

open.substack.comContinual Harness: An Efficient Self-Improving Agent on ARC-AGI-3Continual Harness scores 20.54% on ARC-AGI-3 at $774, showing how reset-free self-improving agents can learn hidden game dynamics at test time.
Seth Karten
@sethkarten.ai

im very glad to see a rebound this year in the total number of NSF GRFP awards to exceed the most this century. Founding the next generation of American scientists is important to keep growth of the sciences.